Hi @dfm,

I noticed that there is a factor of 10 increase in the runtime of the quasisep kernels in the newest version of tinygp. So, I rerun your benchmark notebook in the documentation to test out. And indeed, the quasisep kernel is about 10 times slower than celerite2. I am attaching the reproduced benchmark plot. Note I didn't rerun the benchmark for GPU. Any idea what might be causing it?

I run your notebook on two platforms (M2 Mac and Google Colab), the results are basically consistent. The plot shown was produced using the results obtained with:

• M2 Pro
• Python 3.11.9
• tingpy v0.3

Thanks for your attention!